# initial transient
x = random.rand(40)
for iStep in range(1000):
x = step(x, s)
# primal
history = []
for iChunk in range(nChunks):
history.append([])
for iStep in range(nStepsPerChunk):
history[iChunk].append(x)
x = step(x, s)
# adjoint
nHomo = 20
lss = NILSS(nHomo, x.size, dot)
y = random.rand(nHomo + 1, x.size)
for iChunk in reversed(range(nChunks)):
grad = zeros(y.shape[0])
for iStep in reversed(range(nStepsPerChunk)):
x = history[iChunk][iStep]
for iAdj in range(y.shape[0]):
grad[iAdj] += gradContribution(x, y[iAdj])
y[iAdj] = adjoint(x, y[iAdj], s)
timeFraction = (iStep + iChunk * nStepsPerChunk) / nTotalSteps
windowFunction = sin(timeFraction * pi)**2 * 2
y[-1] += windowFunction / nTotalSteps # inhomogeneous
lss.checkpoint(y, grad)
print(lss.grad())
# initial transient
x = ones(3)
for iStep in range(1000):
x = step(x, s)
# primal
history = []
for iChunk in range(nChunks):
history.append([])
for iStep in range(nStepsPerChunk):
history[iChunk].append(x)
x = step(x, s)
# adjoint
nHomo = 2
lss = NILSS(nHomo, x.size, dot)
y = zeros([nHomo + 1, x.size])
y[0,0] = 1
y[1,1] = 1
for iChunk in reversed(range(nChunks)):
grad = zeros(y.shape[0])
yHist = []
for iStep in reversed(range(nStepsPerChunk)):
x = history[iChunk][iStep]
for iAdj in range(y.shape[0]):
grad[iAdj] += gradContribution(x, y[iAdj])
y[iAdj] = adjoint(x, y[iAdj], s)
timeFraction = (iStep + iChunk * nStepsPerChunk) / nTotalSteps
windowFunction = sin(timeFraction * pi)**2 * 2